yolov3网络结构笔记
2024-04-06 09:53:50
从yolo3_darknet53_coco(gluoncv/model_zoo/yolo/yolo3.py )调试获得，供参考
--------------------------------------------------------------------------------Layer (type) Output Shape Param #
================================================================================Input (1, 3, 512, 703) 0Conv2D-1 (1, 32, 512, 703) 864BatchNorm-2 (1, 32, 512, 703) 128LeakyReLU-3 (1, 32, 512, 703) 0Conv2D-4 (1, 64, 256, 352) 18432BatchNorm-5 (1, 64, 256, 352) 256LeakyReLU-6 (1, 64, 256, 352) 0Conv2D-7 (1, 32, 256, 352) 2048BatchNorm-8 (1, 32, 256, 352) 128LeakyReLU-9 (1, 32, 256, 352) 0Conv2D-10 (1, 64, 256, 352) 18432BatchNorm-11 (1, 64, 256, 352) 256LeakyReLU-12 (1, 64, 256, 352) 0
DarknetBasicBlockV3-13 (1, 64, 256, 352) 0Conv2D-14 (1, 128, 128, 176) 73728BatchNorm-15 (1, 128, 128, 176) 512LeakyReLU-16 (1, 128, 128, 176) 0Conv2D-17 (1, 64, 128, 176) 8192BatchNorm-18 (1, 64, 128, 176) 256LeakyReLU-19 (1, 64, 128, 176) 0Conv2D-20 (1, 128, 128, 176) 73728BatchNorm-21 (1, 128, 128, 176) 512LeakyReLU-22 (1, 128, 128, 176) 0
DarknetBasicBlockV3-23 (1, 128, 128, 176) 0Conv2D-24 (1, 64, 128, 176) 8192BatchNorm-25 (1, 64, 128, 176) 256LeakyReLU-26 (1, 64, 128, 176) 0Conv2D-27 (1, 128, 128, 176) 73728BatchNorm-28 (1, 128, 128, 176) 512LeakyReLU-29 (1, 128, 128, 176) 0
DarknetBasicBlockV3-30 (1, 128, 128, 176) 0Conv2D-31 (1, 256, 64, 88) 294912BatchNorm-32 (1, 256, 64, 88) 1024LeakyReLU-33 (1, 256, 64, 88) 0Conv2D-34 (1, 128, 64, 88) 32768BatchNorm-35 (1, 128, 64, 88) 512LeakyReLU-36 (1, 128, 64, 88) 0Conv2D-37 (1, 256, 64, 88) 294912BatchNorm-38 (1, 256, 64, 88) 1024LeakyReLU-39 (1, 256, 64, 88) 0
DarknetBasicBlockV3-40 (1, 256, 64, 88) 0Conv2D-41 (1, 128, 64, 88) 32768BatchNorm-42 (1, 128, 64, 88) 512LeakyReLU-43 (1, 128, 64, 88) 0Conv2D-44 (1, 256, 64, 88) 294912BatchNorm-45 (1, 256, 64, 88) 1024LeakyReLU-46 (1, 256, 64, 88) 0
DarknetBasicBlockV3-47 (1, 256, 64, 88) 0Conv2D-48 (1, 128, 64, 88) 32768BatchNorm-49 (1, 128, 64, 88) 512LeakyReLU-50 (1, 128, 64, 88) 0Conv2D-51 (1, 256, 64, 88) 294912BatchNorm-52 (1, 256, 64, 88) 1024LeakyReLU-53 (1, 256, 64, 88) 0
DarknetBasicBlockV3-54 (1, 256, 64, 88) 0Conv2D-55 (1, 128, 64, 88) 32768BatchNorm-56 (1, 128, 64, 88) 512LeakyReLU-57 (1, 128, 64, 88) 0Conv2D-58 (1, 256, 64, 88) 294912BatchNorm-59 (1, 256, 64, 88) 1024LeakyReLU-60 (1, 256, 64, 88) 0
DarknetBasicBlockV3-61 (1, 256, 64, 88) 0Conv2D-62 (1, 128, 64, 88) 32768BatchNorm-63 (1, 128, 64, 88) 512LeakyReLU-64 (1, 128, 64, 88) 0Conv2D-65 (1, 256, 64, 88) 294912BatchNorm-66 (1, 256, 64, 88) 1024LeakyReLU-67 (1, 256, 64, 88) 0
DarknetBasicBlockV3-68 (1, 256, 64, 88) 0Conv2D-69 (1, 128, 64, 88) 32768BatchNorm-70 (1, 128, 64, 88) 512LeakyReLU-71 (1, 128, 64, 88) 0Conv2D-72 (1, 256, 64, 88) 294912BatchNorm-73 (1, 256, 64, 88) 1024LeakyReLU-74 (1, 256, 64, 88) 0
DarknetBasicBlockV3-75 (1, 256, 64, 88) 0Conv2D-76 (1, 128, 64, 88) 32768BatchNorm-77 (1, 128, 64, 88) 512LeakyReLU-78 (1, 128, 64, 88) 0Conv2D-79 (1, 256, 64, 88) 294912BatchNorm-80 (1, 256, 64, 88) 1024LeakyReLU-81 (1, 256, 64, 88) 0
DarknetBasicBlockV3-82 (1, 256, 64, 88) 0Conv2D-83 (1, 128, 64, 88) 32768BatchNorm-84 (1, 128, 64, 88) 512LeakyReLU-85 (1, 128, 64, 88) 0Conv2D-86 (1, 256, 64, 88) 294912BatchNorm-87 (1, 256, 64, 88) 1024LeakyReLU-88 (1, 256, 64, 88) 0
DarknetBasicBlockV3-89 (1, 256, 64, 88) 0
darknet53.features[:15]
--------------------------------------------------------------------------------------------------Conv2D-90 (1, 512, 32, 44) 1179648BatchNorm-91 (1, 512, 32, 44) 2048LeakyReLU-92 (1, 512, 32, 44) 0Conv2D-93 (1, 256, 32, 44) 131072BatchNorm-94 (1, 256, 32, 44) 1024LeakyReLU-95 (1, 256, 32, 44) 0Conv2D-96 (1, 512, 32, 44) 1179648BatchNorm-97 (1, 512, 32, 44) 2048LeakyReLU-98 (1, 512, 32, 44) 0
DarknetBasicBlockV3-99 (1, 512, 32, 44) 0Conv2D-100 (1, 256, 32, 44) 131072BatchNorm-101 (1, 256, 32, 44) 1024LeakyReLU-102 (1, 256, 32, 44) 0Conv2D-103 (1, 512, 32, 44) 1179648BatchNorm-104 (1, 512, 32, 44) 2048LeakyReLU-105 (1, 512, 32, 44) 0
DarknetBasicBlockV3-106 (1, 512, 32, 44) 0Conv2D-107 (1, 256, 32, 44) 131072BatchNorm-108 (1, 256, 32, 44) 1024LeakyReLU-109 (1, 256, 32, 44) 0Conv2D-110 (1, 512, 32, 44) 1179648BatchNorm-111 (1, 512, 32, 44) 2048LeakyReLU-112 (1, 512, 32, 44) 0
DarknetBasicBlockV3-113 (1, 512, 32, 44) 0Conv2D-114 (1, 256, 32, 44) 131072BatchNorm-115 (1, 256, 32, 44) 1024LeakyReLU-116 (1, 256, 32, 44) 0Conv2D-117 (1, 512, 32, 44) 1179648BatchNorm-118 (1, 512, 32, 44) 2048LeakyReLU-119 (1, 512, 32, 44) 0
DarknetBasicBlockV3-120 (1, 512, 32, 44) 0Conv2D-121 (1, 256, 32, 44) 131072BatchNorm-122 (1, 256, 32, 44) 1024LeakyReLU-123 (1, 256, 32, 44) 0Conv2D-124 (1, 512, 32, 44) 1179648BatchNorm-125 (1, 512, 32, 44) 2048LeakyReLU-126 (1, 512, 32, 44) 0
DarknetBasicBlockV3-127 (1, 512, 32, 44) 0Conv2D-128 (1, 256, 32, 44) 131072BatchNorm-129 (1, 256, 32, 44) 1024LeakyReLU-130 (1, 256, 32, 44) 0Conv2D-131 (1, 512, 32, 44) 1179648BatchNorm-132 (1, 512, 32, 44) 2048LeakyReLU-133 (1, 512, 32, 44) 0
DarknetBasicBlockV3-134 (1, 512, 32, 44) 0Conv2D-135 (1, 256, 32, 44) 131072BatchNorm-136 (1, 256, 32, 44) 1024LeakyReLU-137 (1, 256, 32, 44) 0Conv2D-138 (1, 512, 32, 44) 1179648BatchNorm-139 (1, 512, 32, 44) 2048LeakyReLU-140 (1, 512, 32, 44) 0
DarknetBasicBlockV3-141 (1, 512, 32, 44) 0Conv2D-142 (1, 256, 32, 44) 131072BatchNorm-143 (1, 256, 32, 44) 1024LeakyReLU-144 (1, 256, 32, 44) 0Conv2D-145 (1, 512, 32, 44) 1179648BatchNorm-146 (1, 512, 32, 44) 2048LeakyReLU-147 (1, 512, 32, 44) 0
DarknetBasicBlockV3-148 (1, 512, 32, 44) 0
darknet53.features[15:24]
--------------------------------------------------------------------------------------------------Conv2D-149 (1, 1024, 16, 22) 4718592BatchNorm-150 (1, 1024, 16, 22) 4096LeakyReLU-151 (1, 1024, 16, 22) 0Conv2D-152 (1, 512, 16, 22) 524288BatchNorm-153 (1, 512, 16, 22) 2048LeakyReLU-154 (1, 512, 16, 22) 0Conv2D-155 (1, 1024, 16, 22) 4718592BatchNorm-156 (1, 1024, 16, 22) 4096LeakyReLU-157 (1, 1024, 16, 22) 0
DarknetBasicBlockV3-158 (1, 1024, 16, 22) 0Conv2D-159 (1, 512, 16, 22) 524288BatchNorm-160 (1, 512, 16, 22) 2048LeakyReLU-161 (1, 512, 16, 22) 0Conv2D-162 (1, 1024, 16, 22) 4718592BatchNorm-163 (1, 1024, 16, 22) 4096LeakyReLU-164 (1, 1024, 16, 22) 0
DarknetBasicBlockV3-165 (1, 1024, 16, 22) 0Conv2D-166 (1, 512, 16, 22) 524288BatchNorm-167 (1, 512, 16, 22) 2048LeakyReLU-168 (1, 512, 16, 22) 0Conv2D-169 (1, 1024, 16, 22) 4718592BatchNorm-170 (1, 1024, 16, 22) 4096LeakyReLU-171 (1, 1024, 16, 22) 0
DarknetBasicBlockV3-172 (1, 1024, 16, 22) 0Conv2D-173 (1, 512, 16, 22) 524288BatchNorm-174 (1, 512, 16, 22) 2048LeakyReLU-175 (1, 512, 16, 22) 0Conv2D-176 (1, 1024, 16, 22) 4718592BatchNorm-177 (1, 1024, 16, 22) 4096LeakyReLU-178 (1, 1024, 16, 22) 0
DarknetBasicBlockV3-179 (1, 1024, 16, 22) 0
darknet53.features[24:]
--------------------------------------------------------------------------------------------------Conv2D-180 (1, 512, 16, 22) 524288BatchNorm-181 (1, 512, 16, 22) 2048LeakyReLU-182 (1, 512, 16, 22) 0Conv2D-183 (1, 1024, 16, 22) 4718592BatchNorm-184 (1, 1024, 16, 22) 4096LeakyReLU-185 (1, 1024, 16, 22) 0Conv2D-186 (1, 512, 16, 22) 524288BatchNorm-187 (1, 512, 16, 22) 2048LeakyReLU-188 (1, 512, 16, 22) 0Conv2D-189 (1, 1024, 16, 22) 4718592BatchNorm-190 (1, 1024, 16, 22) 4096LeakyReLU-191 (1, 1024, 16, 22) 0Conv2D-192 (1, 512, 16, 22) 524288BatchNorm-193 (1, 512, 16, 22) 2048LeakyReLU-194 (1, 512, 16, 22) 0
YOLODetectionBlockV3.body
--------------------------------------------------------------------------------------------------Conv2D-195 (1, 1024, 16, 22) 4718592BatchNorm-196 (1, 1024, 16, 22) 4096LeakyReLU-197 (1, 1024, 16, 22) 0
YOLODetectionBlockV3.tip
--------------------------------------------------------------------------------------------------
YOLODetectionBlockV3-198 (1, 512, 16, 22), (1, 1024, 16, 22) 0
--------------------------------------------------------------------------------------------------Conv2D-199 (1, 255, 16, 22) 261375
YOLOOutputV3.prediction
--------------------------------------------------------------------------------------------------YOLOOutputV3-200 (1, 84480, 6) 32774
YOLOOutputV3
--------------------------------------------------------------------------------------------------Conv2D-201 (1, 256, 16, 22) 131072BatchNorm-202 (1, 256, 16, 22) 1024LeakyReLU-203 (1, 256, 16, 22) 0
YOLOV3.transitions[0]----->_conv2d
--------------------------------------------------------------------------------------------------Conv2D-204 (1, 256, 32, 44) 196608BatchNorm-205 (1, 256, 32, 44) 1024LeakyReLU-206 (1, 256, 32, 44) 0Conv2D-207 (1, 512, 32, 44) 1179648BatchNorm-208 (1, 512, 32, 44) 2048LeakyReLU-209 (1, 512, 32, 44) 0Conv2D-210 (1, 256, 32, 44) 131072BatchNorm-211 (1, 256, 32, 44) 1024LeakyReLU-212 (1, 256, 32, 44) 0Conv2D-213 (1, 512, 32, 44) 1179648BatchNorm-214 (1, 512, 32, 44) 2048LeakyReLU-215 (1, 512, 32, 44) 0Conv2D-216 (1, 256, 32, 44) 131072BatchNorm-217 (1, 256, 32, 44) 1024LeakyReLU-218 (1, 256, 32, 44) 0
YOLODetectionBlockV3.body
--------------------------------------------------------------------------------------------------Conv2D-219 (1, 512, 32, 44) 1179648BatchNorm-220 (1, 512, 32, 44) 2048LeakyReLU-221 (1, 512, 32, 44) 0
YOLODetectionBlockV3.tip
--------------------------------------------------------------------------------------------------
YOLODetectionBlockV3-222 (1, 256, 32, 44), (1, 512, 32, 44) 0
--------------------------------------------------------------------------------------------------Conv2D-223 (1, 255, 32, 44) 130815
YOLOOutputV3.prediction
--------------------------------------------------------------------------------------------------YOLOOutputV3-224 (1, 337920, 6) 32774
YOLOOutputV3
--------------------------------------------------------------------------------------------------Conv2D-225 (1, 128, 32, 44) 32768BatchNorm-226 (1, 128, 32, 44) 512LeakyReLU-227 (1, 128, 32, 44) 0
YOLOV3.transitions[1]----->_conv2d
--------------------------------------------------------------------------------------------------Conv2D-228 (1, 128, 64, 88) 49152BatchNorm-229 (1, 128, 64, 88) 512LeakyReLU-230 (1, 128, 64, 88) 0Conv2D-231 (1, 256, 64, 88) 294912BatchNorm-232 (1, 256, 64, 88) 1024LeakyReLU-233 (1, 256, 64, 88) 0Conv2D-234 (1, 128, 64, 88) 32768BatchNorm-235 (1, 128, 64, 88) 512LeakyReLU-236 (1, 128, 64, 88) 0Conv2D-237 (1, 256, 64, 88) 294912BatchNorm-238 (1, 256, 64, 88) 1024LeakyReLU-239 (1, 256, 64, 88) 0Conv2D-240 (1, 128, 64, 88) 32768BatchNorm-241 (1, 128, 64, 88) 512LeakyReLU-242 (1, 128, 64, 88) 0
YOLODetectionBlockV3.body
--------------------------------------------------------------------------------------------------Conv2D-243 (1, 256, 64, 88) 294912BatchNorm-244 (1, 256, 64, 88) 1024LeakyReLU-245 (1, 256, 64, 88) 0
YOLODetectionBlockV3.tip
--------------------------------------------------------------------------------------------------
YOLODetectionBlockV3-246 (1, 128, 64, 88), (1, 256, 64, 88) 0
--------------------------------------------------------------------------------------------------Conv2D-247 (1, 255, 64, 88) 65535
YOLOOutputV3.prediction
--------------------------------------------------------------------------------------------------YOLOOutputV3-248 (1, 1351680, 6) 32774
YOLOOutputV3
--------------------------------------------------------------------------------------------------YOLOV3-249 (1, 100, 1), (1, 100, 1), (1, 100, 4) 0
================================================================================
打印日志如下，有用的没用的都放这里了
darknet_version = v3
num_layers = 53
darknet layers = [1, 2, 8, 8, 4]
darknet channels = [32, 64, 128, 256, 512, 1024]
anchors = [116. 90. 156. 198. 373. 326.]
self._num_pred = 85
self._num_anchors = 3
all_pred = 255
anchors = [[[[116. 90.][156. 198.][373. 326.]]]]
self.anchors =
[[[[116. 90.][156. 198.][373. 326.]]]]
<NDArray 1x1x3x2 @cpu(0)>
anchors = [ 30. 61. 62. 45. 59. 119.]
self._num_pred = 85
self._num_anchors = 3
all_pred = 255
anchors = [[[[ 30. 61.][ 62. 45.][ 59. 119.]]]]
self.anchors =
[[[[ 30. 61.][ 62. 45.][ 59. 119.]]]]
<NDArray 1x1x3x2 @cpu(0)>
anchors = [10. 13. 16. 30. 33. 23.]
self._num_pred = 85
self._num_anchors = 3
all_pred = 255
anchors = [[[[10. 13.][16. 30.][33. 23.]]]]
self.anchors =
[[[[10. 13.][16. 30.][33. 23.]]]]
<NDArray 1x1x3x2 @cpu(0)>
Shape of pre-processed image: (1, 3, 512, 703)
x.shape = (1, 256, 64, 88)
x.shape = (1, 512, 32, 44)
x.shape = (1, 1024, 16, 22)
pred.shape = (1, 255, 16, 22)
pred.shape = (1, 255, 352)
pred.shape = (1, 352, 3, 85)
raw_box_centers.shape = (1, 352, 3, 2)
raw_box_scales.shape = (1, 352, 3, 2)
objness.shape = (1, 352, 3, 1)
class_pred.shape = (1, 352, 3, 80)
offsets.shape = (1, 1, 128, 128, 2)
slice_like offsets.shape = (1, 1, 16, 22, 2)
reshape offsets.shape = (1, 352, 1, 2)
box_centers.shape = (1, 352, 3, 2)
anchors.shape = (1, 1, 3, 2)
box_scales.shape = (1, 352, 3, 2)
confidence.shape = (1, 352, 3, 1)
class_score.shape = (1, 352, 3, 80)
bbox.shape = (1, 352, 3, 4)
bboxes.shape = (80, 1, 352, 3, 4)
scores.shape = (80, 1, 352, 3, 1)
ids.shape = (80, 1, 352, 3, 1)
detections = (80, 1, 352, 3, 6)
detections = (1, 84480, 6)
dets.shape = (1, 84480, 6)
self.transitions = <class 'mxnet.gluon.nn.basic_layers.HybridSequential'>
x.shape = (1, 256, 16, 22)
upsample.shape = (1, 256, 32, 44)
route_now.shape = (1, 512, 32, 44)
x.shape = (1, 768, 32, 44)
pred.shape = (1, 255, 32, 44)
pred.shape = (1, 255, 1408)
pred.shape = (1, 1408, 3, 85)
raw_box_centers.shape = (1, 1408, 3, 2)
raw_box_scales.shape = (1, 1408, 3, 2)
objness.shape = (1, 1408, 3, 1)
class_pred.shape = (1, 1408, 3, 80)
offsets.shape = (1, 1, 128, 128, 2)
slice_like offsets.shape = (1, 1, 32, 44, 2)
reshape offsets.shape = (1, 1408, 1, 2)
box_centers.shape = (1, 1408, 3, 2)
anchors.shape = (1, 1, 3, 2)
box_scales.shape = (1, 1408, 3, 2)
confidence.shape = (1, 1408, 3, 1)
class_score.shape = (1, 1408, 3, 80)
bbox.shape = (1, 1408, 3, 4)
bboxes.shape = (80, 1, 1408, 3, 4)
scores.shape = (80, 1, 1408, 3, 1)
ids.shape = (80, 1, 1408, 3, 1)
detections = (80, 1, 1408, 3, 6)
detections = (1, 337920, 6)
dets.shape = (1, 337920, 6)
self.transitions = <class 'mxnet.gluon.nn.basic_layers.HybridSequential'>
x.shape = (1, 128, 32, 44)
upsample.shape = (1, 128, 64, 88)
route_now.shape = (1, 256, 64, 88)
x.shape = (1, 384, 64, 88)
pred.shape = (1, 255, 64, 88)
pred.shape = (1, 255, 5632)
pred.shape = (1, 5632, 3, 85)
raw_box_centers.shape = (1, 5632, 3, 2)
raw_box_scales.shape = (1, 5632, 3, 2)
objness.shape = (1, 5632, 3, 1)
class_pred.shape = (1, 5632, 3, 80)
offsets.shape = (1, 1, 128, 128, 2)
slice_like offsets.shape = (1, 1, 64, 88, 2)
reshape offsets.shape = (1, 5632, 1, 2)
box_centers.shape = (1, 5632, 3, 2)
anchors.shape = (1, 1, 3, 2)
box_scales.shape = (1, 5632, 3, 2)
confidence.shape = (1, 5632, 3, 1)
class_score.shape = (1, 5632, 3, 80)
bbox.shape = (1, 5632, 3, 4)
bboxes.shape = (80, 1, 5632, 3, 4)
scores.shape = (80, 1, 5632, 3, 1)
ids.shape = (80, 1, 5632, 3, 1)
detections = (80, 1, 5632, 3, 6)
detections = (1, 1351680, 6)
dets.shape = (1, 1351680, 6)
all_detections.shape = 3
all_detections[0].shape = (1, 84480, 6)
result.shape = (1, 1774080, 6)
after result.shape = (1, 1774080, 6)
slice result.shape = (1, 100, 6)
自己总结的计算过程如下：
img ——> darknet53 ——> YOLODetectionBlockV3.body ——> YOLODetectionBlockV3.tip ——> YOLOOutputV3.prediction ——>
(reshape, transpose, slice_axis) ——> [raw_box_centers(2), raw_box_scales(2), objness(1), class_pred(80)] ——>
[box_centers(offsets, _stride), box_scales(anchors), confidence, class_score(class_pred * confidence)] ——>
[bboxes(4), scores(1)] + ids(from arrange,1) ——> (reshape) ——> detections ——> box_nms ——> slice_axis ——>
(ids, scores, bboxes)class YOLOV3(gluon.HybridBlock): 添加的打印如下：
def hybrid_forward(self, F, x, *args):"""YOLOV3 network hybrid forward.Parameters----------F : mxnet.nd or mxnet.sym`F` is mxnet.sym if hybridized or mxnet.nd if not.x : mxnet.nd.NDArrayInput data.*args : optional, mxnet.nd.NDArrayDuring training, extra inputs are required:(gt_boxes, obj_t, centers_t, scales_t, weights_t, clas_t)These are generated by YOLOV3PrefetchTargetGenerator in dataloader transform function.Returns-------(tuple of) mxnet.nd.NDArrayDuring inference, return detections in shape (B, N, 6)with format (cid, score, xmin, ymin, xmax, ymax)During training, return losses only: (obj_loss, center_loss, scale_loss, cls_loss)."""all_box_centers = []all_box_scales = []all_objectness = []all_class_pred = []all_anchors = []all_offsets = []all_feat_maps = []all_detections = []routes = []for stage in self.stages:x = stage(x)routes.append(x)print("x.shape = ", x.shape)# the YOLO output layers are used in reverse order, i.e., from very deep layers to shallowfor i, block, output in zip(range(len(routes)), self.yolo_blocks, self.yolo_outputs):x, tip = block(x)dets = output(tip)print("dets.shape = ", dets.shape)all_detections.append(dets)if i >= len(routes) - 1:break# add transition layersprint("self.transitions = ", type(self.transitions))x = self.transitions[i](x)print("x.shape = ", x.shape)# upsample feature map reverse to shallow layersupsample = _upsample(x, stride=2)print("upsample.shape = ", upsample.shape)route_now = routes[::-1][i + 1]print("route_now.shape = ", route_now.shape)x = F.concat(upsample, route_now, dim=1)print("x.shape = ", x.shape)print("all_detections.shape = ", len(all_detections))print("all_detections[0].shape = ", all_detections[0].shape)# concat all detection results from different stagesresult = F.concat(*all_detections, dim=1)print("result.shape = ", result.shape)# apply nms per classif self.nms_thresh > 0 and self.nms_thresh < 1:result = F.contrib.box_nms(result, overlap_thresh=self.nms_thresh, valid_thresh=0.01,topk=self.nms_topk, id_index=0, score_index=1, coord_start=2, force_suppress=False)print("after result.shape = ", result.shape)if self.post_nms > 0:result = result.slice_axis(axis=1, begin=0, end=self.post_nms)print("slice result.shape = ", result.shape)ids = result.slice_axis(axis=-1, begin=0, end=1)scores = result.slice_axis(axis=-1, begin=1, end=2)bboxes = result.slice_axis(axis=-1, begin=2, end=None)return ids, scores, bboxes
class YOLOOutputV3(gluon.HybridBlock):添加的打印如下：
class YOLOOutputV3(gluon.HybridBlock):"""YOLO output layer V3.Parameters----------index : intIndex of the yolo output layer, to avoid naming conflicts only.num_class : intNumber of foreground objects.anchors : iterableThe anchor setting. Reference: https://arxiv.org/pdf/1804.02767.pdf.stride : intStride of feature map.alloc_size : tuple of int, default is (128, 128)For advanced users. Define `alloc_size` to generate large enough anchormaps, which will later saved in parameters. During inference, we support arbitraryinput image by cropping corresponding area of the anchor map. This allow usto export to symbol so we can run it in c++, Scalar, etc."""def __init__(self, index, num_class, anchors, stride,alloc_size=(128, 128), **kwargs):super(YOLOOutputV3, self).__init__(**kwargs)anchors = np.array(anchors).astype('float32')print("anchors = ", anchors)self._classes = num_classself._num_pred = 1 + 4 + num_class # 1 objness + 4 box + num_classself._num_anchors = anchors.size // 2self._stride = stridewith self.name_scope():print("self._num_pred = ", self._num_pred)print("self._num_anchors = ", self._num_anchors)all_pred = self._num_pred * self._num_anchorsprint("all_pred = %d" % all_pred)self.prediction = nn.Conv2D(all_pred, kernel_size=1, padding=0, strides=1)# anchors will be multiplied to predictionsanchors = anchors.reshape(1, 1, -1, 2)print("anchors = ", anchors)self.anchors = self.params.get_constant('anchor_%d'%(index), anchors)print("self.anchors = ", self.anchors.value)# offsets will be added to predictionsgrid_x = np.arange(alloc_size[1])grid_y = np.arange(alloc_size[0])grid_x, grid_y = np.meshgrid(grid_x, grid_y)# stack to (n, n, 2)offsets = np.concatenate((grid_x[:, :, np.newaxis], grid_y[:, :, np.newaxis]), axis=-1)# expand dims to (1, 1, n, n, 2) so it's easier for broadcastingoffsets = np.expand_dims(np.expand_dims(offsets, axis=0), axis=0)self.offsets = self.params.get_constant('offset_%d'%(index), offsets)def hybrid_forward(self, F, x, anchors, offsets):"""Hybrid Forward of YOLOV3Output layer.Parameters----------F : mxnet.nd or mxnet.sym`F` is mxnet.sym if hybridized or mxnet.nd if not.x : mxnet.nd.NDArrayInput feature map.anchors : mxnet.nd.NDArrayAnchors loaded from self, no need to supply.offsets : mxnet.nd.NDArrayOffsets loaded from self, no need to supply.Returns-------(tuple of) mxnet.nd.NDArrayDuring training, return (bbox, raw_box_centers, raw_box_scales, objness,class_pred, anchors, offsets).During inference, return detections."""# prediction flat to (batch, pred per pixel, height * width)pred = self.prediction(x)print("pred.shape = ", pred.shape)pred = pred.reshape((0, self._num_anchors * self._num_pred, -1))print("pred.shape = ", pred.shape)# transpose to (batch, height * width, num_anchor, num_pred)pred = pred.transpose(axes=(0, 2, 1)).reshape((0, -1, self._num_anchors, self._num_pred))print("pred.shape = ", pred.shape)# componentsraw_box_centers = pred.slice_axis(axis=-1, begin=0, end=2)print("raw_box_centers.shape = ", raw_box_centers.shape)raw_box_scales = pred.slice_axis(axis=-1, begin=2, end=4)print("raw_box_scales.shape = ", raw_box_scales.shape)objness = pred.slice_axis(axis=-1, begin=4, end=5)print("objness.shape = ", objness.shape)class_pred = pred.slice_axis(axis=-1, begin=5, end=None)print("class_pred.shape = ", class_pred.shape)# valid offsets, (1, 1, height, width, 2)print("offsets.shape = ", offsets.shape)offsets = F.slice_like(offsets, x * 0, axes=(2, 3))print("slice_like offsets.shape = ", offsets.shape)# reshape to (1, height*width, 1, 2)offsets = offsets.reshape((1, -1, 1, 2))print("reshape offsets.shape = ", offsets.shape)# print(offsets)box_centers = F.broadcast_add(F.sigmoid(raw_box_centers), offsets) * self._strideprint("box_centers.shape = ", box_centers.shape)print("anchors.shape = ", anchors.shape)box_scales = F.broadcast_mul(F.exp(raw_box_scales), anchors)print("box_scales.shape = ", box_scales.shape)confidence = F.sigmoid(objness)print("confidence.shape = ", confidence.shape)class_score = F.broadcast_mul(F.sigmoid(class_pred), confidence)print("class_score.shape = ", class_score.shape)wh = box_scales / 2.0bbox = F.concat(box_centers - wh, box_centers + wh, dim=-1)print("bbox.shape = ", bbox.shape)# prediction per classbboxes = F.tile(bbox, reps=(self._classes, 1, 1, 1, 1))print("bboxes.shape = ", bboxes.shape)scores = F.transpose(class_score, axes=(3, 0, 1, 2)).expand_dims(axis=-1)print("scores.shape = ", scores.shape)ids = F.broadcast_add(scores * 0, F.arange(0, self._classes).reshape((0, 1, 1, 1, 1)))print("ids.shape = ", ids.shape)detections = F.concat(ids, scores, bboxes, dim=-1)print("detections = ", detections.shape)# reshape to (B, xx, 6)detections = F.reshape(detections.transpose(axes=(1, 0, 2, 3, 4)), (0, -1, 6))print("detections = ", detections.shape)return detectionsyolov3网络结构笔记相关推荐
- YOLOv3 学习笔记：大神好贴汇总+自身经验记录
一.学习 YOLOv3 YOLOv3论文 YOLOv3:An Incremental Improvement 全文翻译 [目标检测简史]进击的 YOLOv3,目标检测网络的巅峰之作 语言生动易懂,适合 ...
- 【华为云技术分享】小白学YOLO：YOLOv3网络结构细致解析
Yolov3 网络结构 在博客"Yolo发展历史及网络结构"中我们已经详细的解释了Yolov1的网络结构,并简要的提到了Yolov2与Yolov3对于网络结构的改进,本篇博客将详细 ...
- YOLOv3网络结构细致解析
YOLOv3网络结构细致解析 先贴一张结构图镇楼: layer filters size input output 0 conv 32 3 x 3 / 1 416 x 416 x 3 -> 41 ...
- YoloV3学习笔记（一）—— 打标签
YoloV3学习笔记(一) labelimg标注图片 labelimg页面介绍 报错分析 解决方法 labelimg标注图片 搭建好Yolov3+pytorch+windows环境后,为了训练模型,标 ...
- YOLOV3学习笔记（噼里啪啦up主）
YOLOV3学习笔记(噼里啪啦up主) YOLOV3上面的IOU LOSS是错误的,因为在两个候选框和目标框围成的面积相同时,YOLOV3并没有告诉网络应该选哪个候选框. 2. 在YOLOV3中判断候 ...
- pytorch实现Yolov3网络结构
#基于torch的yolov3网络结构Darknet53实现 最近一段时间在研究目标检测,先从yolov3看起.本文讲下如何使用pytorch实现yolov3网络结构中的DarkNet53网络结构. ...
- YOLOV3 网络结构学习笔记
注:本文非原创,文章内容都是引用以下文章中,本文只是记录学习笔记. yolo系列之yolo v3[深度解析]_木盏的博客-CSDN博客_yolo3 YOLO v3算法详解_'Atlas'的博客-CSD ...
- 最详细的YOLOv3论文笔记
个人博客:http://www.chenjianqu.com/ 原文链接:http://www.chenjianqu.com/show-117.html 论文:YOLOv3: An Increment ...
- YOLOV3网络结构搭建
YOLOV3 一.定义一个残差结构 # 残差结构 # 利用一个1x1卷积下降通道数,然后利用一个3x3卷积提取特征并且上升通道数 # 最后接上一个残差边 #---------------------- ...
最新文章
- Android的数据存储方式
- centos7重新加载服务的命令_Centos7 systemctl添加service服务参数说明
- GraphQL报错:error: NetworkError when attempting to fetch resource.
- HDOJ 1047 Integer Inquiry
- html 滤镜兼容,CSS3 中各种滤镜介绍和兼容性说明 - 文章教程
- 产品销售的过程中，价值是不断被定义和挖掘的
- 13.python中web框架概念的引入。
- django中使用第三方包实现定时任务
- 智能优化算法应用：基于GWO优化的灰度熵图像多阈值分割 - 附代码
- ffmpeg drawtext 背景_8款电视背景墙：电视背景墙这样装，不仅省钱还作用多！效果大不一样!...
- vue : 无法加载文件 D:\Program Files\nodejs\vue.ps1，因为在此系统上禁止运行脚本。有关详细信息，请参阅 https:/go.microsof t.com/fwlin
- 人型自走输入法（网页日语输入法）
- 【图像加密】基于LSB算法实现图像隐写含Matlab源码
- android 仿百度地图,仿百度地图街景实现
- 文字竖直居中 html,CSS 文字垂直居中
- java实现手机验证码功能
- python无法调用urlopen_python urlopen 不了一个url 而用curl 是可以的
- mysql 编程算法_十大编程算法助程序员走上高手之路
- STM32使用SWD下载
- .net Core 在 CentOS7下,报The type initializer for ‘Gdip‘ threw an exception.异常